Crystal structure prediction by combining graph network and optimization algorithm.

Area of Science:

• Condensed matter physics
• Chemical science
• Materials science

Background:

• Crystal structure prediction is a critical yet challenging problem in materials science.
• Accurate prediction of crystal structures is essential for discovering new materials with desired properties.

Purpose of the Study:

• To develop and evaluate a novel machine-learning approach for accelerated crystal structure prediction.
• To establish a flexible framework combining materials databases, graph networks, and optimization algorithms.

Main Methods:

• Utilized a machine-learning framework integrating a materials database (OQMD, Matbench), a graph network (GN) model, and an optimization algorithm (OA).
• Compared different optimization algorithms including random searching (RAS), particle-swarm optimization (PSO), and Bayesian optimization (BO).

Main Results:

• The graph network model trained on the Matbench database combined with Bayesian optimization (GN(MatB)-BO) demonstrated superior performance.
• Achieved accurate crystal structure predictions for 29 compounds with a computational cost three orders of magnitude lower than conventional methods.
• The framework proved flexible and adaptable to different databases and optimization strategies.

Conclusions:

• The proposed machine-learning approach significantly accelerates crystal structure prediction.
• This data-driven methodology offers a promising avenue for efficient materials discovery and design.